Feature #2993
Fires on “ print” (not labeled yet: read the examples)
- Fires on
- 8.14% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 95%
- story 3%
- python 1%
- chat 1%
- worksheet 0%
Makes the model say
HilJ seek `"replaceAll answer rain
Makes it not say
!9 upon```ant. remember Mia
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.97module>⏎ print(dercement
-
agent 3.82utils.py", line 5, in <module>⏎ assert half(3) ==
-
agent 3.77", line 3, in <module>⏎ assert max_item([0
-
agent 3.77core.py", line 4, in <module>⏎ assert echo("sun",
-
agent 3.76line 3, in <module>⏎ assert divides(17,
-
agent 3.593, in <module>⏎ assert cubed(10)
-
agent 3.51tools.py", line 3, in <module>⏎ assert raise_to(1
-
agent 3.5calc.py", line 3, in <module>⏎ assert power(4,
-
agent 3.46textlib.py", line 6, in <module>⏎ assert repeat("hello",
-
agent 3.45with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def negate<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 3.41utils.py", line 3, in <module>⏎ assert fact(3) ==
-
agent 3.39<module>⏎ print(posiitive(